This is fun, thanks. But I do have the following thought: If I identify bikes in a bike rack and wait for you to take one, I can watch you going through the city, including your final destination? Mildly scary in some circumstances.

5. If I change something by hand (e.g. replacing \n's by actual newlines in strings), the models tent to turn it back to what they did before.

I am sure there are ways to fix some of these issues, or to lessen their impact. But this is my current state of #VibeCoding.

4. The overall quality of the code is a disaster! And refactoring creates more problems than actually cleaning the code. It really looks like someone programs it and restarts its understanding of what is there and why every hour, see 2.

3. GPT-4.1 makes much smaller changes and implement features much less creative than Claude Sonnet 4. It was often good to first let it come up with a solution completely by itself so that I can see ways of doing it (in particular visual components on the website) I hadn't thought of before.

2. The stateless concept of LLMs is problematic: While I remember what we did, the models do not. They have no idea what features exists, how they are implemented, where in the codebase a new feature is perfectly placed, the naming scheme of variables, or anything.

1. Before editing, discuss the concrete road map for every single change. Otherwise, the models (especially Claude Sonnet 4 and GPT-5) tent to rewrite your code, introducing problems all over the place.